Frameworks for interpretability of deep learning-based prognostics and health management

• Ano de defesa: 2022
• Autor(a) principal: Figueroa Barraza, Joaquín Eduardo
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Contrafactual; Interpretabilidade; Interpretability; Neural networks; Prognóstico e gestão da saúde; Redes neurais; Seleção de variáveis
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/3/3135/tde-22052023-151410/

In the last five years, the use of deep learning algorithms for prognostics and health management (PHM) has led to a performance increase in fault diagnostics, prognostics, and anomaly detection. However, the lack of explanation and interpretability of these models results in a resistance towards their credibility and deployment. This means that even though deep learning-based models may achieve great performance, the understanding and explanation of how a deep learning-based PHM model obtains its results is still an open area of research. In this thesis, three techniques for interpretability of deep learning models in the context of prognostics and health management are proposed. The first one is comprised of a technique for feature selection and a methodology for quantitative evaluation of the techniques performance and comparison with other techniques. The proposed technique consists of a hidden layer next to the input layer whose weights determine the importance of each feature within the model. These weights are trained jointly with the rest of the network. The layer is referred to as feature selection (FS) layer. Moreover, the methodology for evaluation proposes the use of a novel metric referred to as ranking quality score (RQS). For the second framework, a multi-task neural network, referred to as Sparse Counterfactual Generation Neural Network (SCF-Net), is proposed for simultaneous fault diagnosis and counterfactual generation. Thus, the network has the ability to diagnose health states and deliver information referring to the minimal changes in the input values that lead to a change in the predicted health state by the model. In the third framework, the two previous approaches are combined in a network architecture referred to as Feature Selection and Sparse Counterfactual Generation network (FS-SCF). Also, a methodology is proposed for calculation of causality-based values for each feature, such as necessity, sufficiency, (necessity or sufficiency) and (necessity and sufficiency). This is used to further analyze the model and to interpret the results obtained from the FS layer. For these three frameworks, several case studies are used for testing, and compared to other existing techniques. Results across the three frameworks show a successful increase in interpretability while keeping task performance at the same level. Thus, the accuracy/interpretability tradeoff is successfully addressed in this thesis. Future lines of research include testing in other kinds of neural networks, such as convolutional neural networks, recurrent neural networks, and transformers. In the case of counterfactual-based approaches, future works include their adaption for regression tasks, due to the fact that they are limited to classification. This could increase the types of applications in PHM. For example, remaining useful life (RUL) prediction.